Debugging Correctness Issues in Training Machine Learning Models

3. The computer-implemented method of claim 2, wherein the one or more application-specific integrated circuits comprise one or more artificial intelligence accelerators.

4. The computer-implemented method of claim 2, wherein the one or more application-specific integrated circuits comprise one or more tensor processing units.

9. The computer-implemented method of claim 8, wherein the one or more special values comprise one or more not-a-number values.

10. The computer-implemented method of claim 1, wherein the shared training operation comprises at least one of determining a value of an objective function, determining a gradient of the objective function, determining an output of a neural network layer, determining a result of a convolution operation of a neural network layer, or determining a result of an activation function of a neural network layer.

12. The computer-implemented method of claim 11, wherein a corresponding signature for a matrix represents at least one of whether any element of the matrix comprises a special value, a matrix norm of the matrix, or a proper subset of the matrix.

15. The system of claim 14, wherein the one or more application-specific integrated circuits comprise one or more artificial intelligence accelerators.

16. The system of claim 14, wherein the one or more application-specific integrated circuits comprise one or more tensor processing units.

21. The system of claim 20, wherein the one or more special values comprise one or more not-a-number values.

22. The system of claim 13, wherein the shared training operation comprises at least one of determining a value of an objective function, determining a gradient of the objective function, determining an output of a neural network layer, determining a result of a convolution operation of a neural network layer, or determining a result of an activation function of a neural network layer.

24. The system of claim 23, wherein a corresponding signature for a matrix represents at least one of whether any element of the matrix comprises a special value, a matrix norm of the matrix, or a proper subset of the matrix.